blob: 1a6de0a7d8eba53a516bc0f6359cbd5240580b15 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Becky Bruce41151e72011-06-28 09:54:48 +00002 * PPC Huge TLB Page Support for Kernel.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
Becky Bruce41151e72011-06-28 09:54:48 +00005 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/mm.h>
David Gibson883a3e52009-10-26 19:24:31 +000012#include <linux/io.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090013#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/hugetlb.h>
Paul Mackerras342d3db2011-12-12 12:38:05 +000015#include <linux/export.h>
Becky Bruce41151e72011-06-28 09:54:48 +000016#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/bootmem.h>
Kumar Gala13020be2011-11-24 09:40:07 +000019#include <linux/moduleparam.h>
David Gibson883a3e52009-10-26 19:24:31 +000020#include <asm/pgtable.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <asm/pgalloc.h>
22#include <asm/tlb.h>
Becky Bruce41151e72011-06-28 09:54:48 +000023#include <asm/setup.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024
Jon Tollefson91224342008-07-23 21:27:55 -070025#define PAGE_SHIFT_64K 16
26#define PAGE_SHIFT_16M 24
27#define PAGE_SHIFT_16G 34
Jon Tollefson4ec161c2008-01-04 09:59:50 +110028
Becky Bruce41151e72011-06-28 09:54:48 +000029unsigned int HPAGE_SHIFT;
30
31/*
32 * Tracks gpages after the device tree is scanned and before the
Becky Brucea6146882011-10-10 10:50:43 +000033 * huge_boot_pages list is ready. On non-Freescale implementations, this is
34 * just used to track 16G pages and so is a single array. FSL-based
35 * implementations may have more than one gpage size, so we need multiple
36 * arrays
Becky Bruce41151e72011-06-28 09:54:48 +000037 */
Becky Bruce881fde12011-10-10 10:50:40 +000038#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +000039#define MAX_NUMBER_GPAGES 128
40struct psize_gpages {
41 u64 gpage_list[MAX_NUMBER_GPAGES];
42 unsigned int nr_gpages;
43};
44static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
Becky Bruce881fde12011-10-10 10:50:40 +000045#else
46#define MAX_NUMBER_GPAGES 1024
47static u64 gpage_freearray[MAX_NUMBER_GPAGES];
48static unsigned nr_gpages;
Becky Bruce41151e72011-06-28 09:54:48 +000049#endif
David Gibsonf10a04c2006-04-28 15:02:51 +100050
Jon Tollefson0d9ea752008-07-23 21:27:56 -070051static inline int shift_to_mmu_psize(unsigned int shift)
52{
David Gibsond1837cb2009-10-26 19:24:31 +000053 int psize;
54
55 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize)
56 if (mmu_psize_defs[psize].shift == shift)
57 return psize;
Jon Tollefson0d9ea752008-07-23 21:27:56 -070058 return -1;
59}
60
61static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
62{
63 if (mmu_psize_defs[mmu_psize].shift)
64 return mmu_psize_defs[mmu_psize].shift;
65 BUG();
66}
67
David Gibsona4fe3ce2009-10-26 19:24:31 +000068#define hugepd_none(hpd) ((hpd).pd == 0)
69
David Gibsona4fe3ce2009-10-26 19:24:31 +000070pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
David Gibsonf10a04c2006-04-28 15:02:51 +100071{
David Gibsona4fe3ce2009-10-26 19:24:31 +000072 pgd_t *pg;
73 pud_t *pu;
74 pmd_t *pm;
75 hugepd_t *hpdp = NULL;
76 unsigned pdshift = PGDIR_SHIFT;
77
78 if (shift)
79 *shift = 0;
80
81 pg = pgdir + pgd_index(ea);
82 if (is_hugepd(pg)) {
83 hpdp = (hugepd_t *)pg;
84 } else if (!pgd_none(*pg)) {
85 pdshift = PUD_SHIFT;
86 pu = pud_offset(pg, ea);
87 if (is_hugepd(pu))
88 hpdp = (hugepd_t *)pu;
89 else if (!pud_none(*pu)) {
90 pdshift = PMD_SHIFT;
91 pm = pmd_offset(pu, ea);
92 if (is_hugepd(pm))
93 hpdp = (hugepd_t *)pm;
94 else if (!pmd_none(*pm)) {
Becky Bruce41151e72011-06-28 09:54:48 +000095 return pte_offset_kernel(pm, ea);
David Gibsona4fe3ce2009-10-26 19:24:31 +000096 }
97 }
98 }
99
100 if (!hpdp)
101 return NULL;
102
103 if (shift)
104 *shift = hugepd_shift(*hpdp);
105 return hugepte_offset(hpdp, ea, pdshift);
106}
Paul Mackerras342d3db2011-12-12 12:38:05 +0000107EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000108
109pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110{
111 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
112}
113
114static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
115 unsigned long address, unsigned pdshift, unsigned pshift)
116{
Becky Bruce41151e72011-06-28 09:54:48 +0000117 struct kmem_cache *cachep;
118 pte_t *new;
119
Becky Bruce881fde12011-10-10 10:50:40 +0000120#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000121 int i;
122 int num_hugepd = 1 << (pshift - pdshift);
123 cachep = hugepte_cache;
Becky Bruce881fde12011-10-10 10:50:40 +0000124#else
125 cachep = PGT_CACHE(pdshift - pshift);
Becky Bruce41151e72011-06-28 09:54:48 +0000126#endif
127
128 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
David Gibsonf10a04c2006-04-28 15:02:51 +1000129
David Gibsona4fe3ce2009-10-26 19:24:31 +0000130 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
131 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
132
David Gibsonf10a04c2006-04-28 15:02:51 +1000133 if (! new)
134 return -ENOMEM;
135
136 spin_lock(&mm->page_table_lock);
Becky Bruce881fde12011-10-10 10:50:40 +0000137#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000138 /*
139 * We have multiple higher-level entries that point to the same
140 * actual pte location. Fill in each as we go and backtrack on error.
141 * We need all of these so the DTLB pgtable walk code can find the
142 * right higher-level entry without knowing if it's a hugepage or not.
143 */
144 for (i = 0; i < num_hugepd; i++, hpdp++) {
145 if (unlikely(!hugepd_none(*hpdp)))
146 break;
147 else
148 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
149 }
150 /* If we bailed from the for loop early, an error occurred, clean up */
151 if (i < num_hugepd) {
152 for (i = i - 1 ; i >= 0; i--, hpdp--)
153 hpdp->pd = 0;
154 kmem_cache_free(cachep, new);
155 }
Becky Brucea1cd5412011-10-10 10:50:39 +0000156#else
157 if (!hugepd_none(*hpdp))
158 kmem_cache_free(cachep, new);
159 else
160 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
Becky Bruce41151e72011-06-28 09:54:48 +0000161#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000162 spin_unlock(&mm->page_table_lock);
163 return 0;
164}
165
Becky Brucea1cd5412011-10-10 10:50:39 +0000166/*
167 * These macros define how to determine which level of the page table holds
168 * the hpdp.
169 */
170#ifdef CONFIG_PPC_FSL_BOOK3E
171#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
172#define HUGEPD_PUD_SHIFT PUD_SHIFT
173#else
174#define HUGEPD_PGD_SHIFT PUD_SHIFT
175#define HUGEPD_PUD_SHIFT PMD_SHIFT
176#endif
177
David Gibsona4fe3ce2009-10-26 19:24:31 +0000178pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
179{
180 pgd_t *pg;
181 pud_t *pu;
182 pmd_t *pm;
183 hugepd_t *hpdp = NULL;
184 unsigned pshift = __ffs(sz);
185 unsigned pdshift = PGDIR_SHIFT;
David Gibson0b264252008-09-05 11:49:54 +1000186
David Gibsona4fe3ce2009-10-26 19:24:31 +0000187 addr &= ~(sz-1);
188
189 pg = pgd_offset(mm, addr);
Becky Brucea1cd5412011-10-10 10:50:39 +0000190
191 if (pshift >= HUGEPD_PGD_SHIFT) {
David Gibsona4fe3ce2009-10-26 19:24:31 +0000192 hpdp = (hugepd_t *)pg;
193 } else {
194 pdshift = PUD_SHIFT;
195 pu = pud_alloc(mm, pg, addr);
Becky Brucea1cd5412011-10-10 10:50:39 +0000196 if (pshift >= HUGEPD_PUD_SHIFT) {
David Gibsona4fe3ce2009-10-26 19:24:31 +0000197 hpdp = (hugepd_t *)pu;
198 } else {
199 pdshift = PMD_SHIFT;
200 pm = pmd_alloc(mm, pu, addr);
201 hpdp = (hugepd_t *)pm;
202 }
203 }
204
205 if (!hpdp)
206 return NULL;
207
208 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
209
210 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
211 return NULL;
212
213 return hugepte_offset(hpdp, addr, pdshift);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100214}
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100215
Becky Bruce881fde12011-10-10 10:50:40 +0000216#ifdef CONFIG_PPC_FSL_BOOK3E
Jon Tollefson658013e2008-07-23 21:27:54 -0700217/* Build list of addresses of gigantic pages. This function is used in early
218 * boot before the buddy or bootmem allocator is setup.
219 */
Becky Bruce41151e72011-06-28 09:54:48 +0000220void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
221{
222 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
223 int i;
224
225 if (addr == 0)
226 return;
227
228 gpage_freearray[idx].nr_gpages = number_of_pages;
229
230 for (i = 0; i < number_of_pages; i++) {
231 gpage_freearray[idx].gpage_list[i] = addr;
232 addr += page_size;
233 }
234}
235
236/*
237 * Moves the gigantic page addresses from the temporary list to the
238 * huge_boot_pages list.
239 */
240int alloc_bootmem_huge_page(struct hstate *hstate)
241{
242 struct huge_bootmem_page *m;
243 int idx = shift_to_mmu_psize(hstate->order + PAGE_SHIFT);
244 int nr_gpages = gpage_freearray[idx].nr_gpages;
245
246 if (nr_gpages == 0)
247 return 0;
248
249#ifdef CONFIG_HIGHMEM
250 /*
251 * If gpages can be in highmem we can't use the trick of storing the
252 * data structure in the page; allocate space for this
253 */
254 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
255 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
256#else
257 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
258#endif
259
260 list_add(&m->list, &huge_boot_pages);
261 gpage_freearray[idx].nr_gpages = nr_gpages;
262 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
263 m->hstate = hstate;
264
265 return 1;
266}
267/*
268 * Scan the command line hugepagesz= options for gigantic pages; store those in
269 * a list that we use to allocate the memory once all options are parsed.
270 */
271
272unsigned long gpage_npages[MMU_PAGE_COUNT];
273
Paul Gortmaker89528122012-05-07 10:32:22 -0400274static int __init do_gpage_early_setup(char *param, char *val,
275 const char *unused)
Becky Bruce41151e72011-06-28 09:54:48 +0000276{
277 static phys_addr_t size;
278 unsigned long npages;
279
280 /*
281 * The hugepagesz and hugepages cmdline options are interleaved. We
282 * use the size variable to keep track of whether or not this was done
283 * properly and skip over instances where it is incorrect. Other
284 * command-line parsing code will issue warnings, so we don't need to.
285 *
286 */
287 if ((strcmp(param, "default_hugepagesz") == 0) ||
288 (strcmp(param, "hugepagesz") == 0)) {
289 size = memparse(val, NULL);
290 } else if (strcmp(param, "hugepages") == 0) {
291 if (size != 0) {
292 if (sscanf(val, "%lu", &npages) <= 0)
293 npages = 0;
294 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
295 size = 0;
296 }
297 }
298 return 0;
299}
300
301
302/*
303 * This function allocates physical space for pages that are larger than the
304 * buddy allocator can handle. We want to allocate these in highmem because
305 * the amount of lowmem is limited. This means that this function MUST be
306 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
307 * allocate to grab highmem.
308 */
309void __init reserve_hugetlb_gpages(void)
310{
311 static __initdata char cmdline[COMMAND_LINE_SIZE];
312 phys_addr_t size, base;
313 int i;
314
315 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
Pawel Moll026cee02012-03-26 12:50:51 +1030316 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
317 &do_gpage_early_setup);
Becky Bruce41151e72011-06-28 09:54:48 +0000318
319 /*
320 * Walk gpage list in reverse, allocating larger page sizes first.
321 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
322 * When we reach the point in the list where pages are no longer
323 * considered gpages, we're done.
324 */
325 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
326 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
327 continue;
328 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
329 break;
330
331 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
332 base = memblock_alloc_base(size * gpage_npages[i], size,
333 MEMBLOCK_ALLOC_ANYWHERE);
334 add_gpage(base, size, gpage_npages[i]);
335 }
336}
337
Becky Bruce881fde12011-10-10 10:50:40 +0000338#else /* !PPC_FSL_BOOK3E */
Becky Bruce41151e72011-06-28 09:54:48 +0000339
340/* Build list of addresses of gigantic pages. This function is used in early
341 * boot before the buddy or bootmem allocator is setup.
342 */
343void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
Jon Tollefson658013e2008-07-23 21:27:54 -0700344{
345 if (!addr)
346 return;
347 while (number_of_pages > 0) {
348 gpage_freearray[nr_gpages] = addr;
349 nr_gpages++;
350 number_of_pages--;
351 addr += page_size;
352 }
353}
354
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700355/* Moves the gigantic page addresses from the temporary list to the
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700356 * huge_boot_pages list.
357 */
358int alloc_bootmem_huge_page(struct hstate *hstate)
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700359{
360 struct huge_bootmem_page *m;
361 if (nr_gpages == 0)
362 return 0;
363 m = phys_to_virt(gpage_freearray[--nr_gpages]);
364 gpage_freearray[nr_gpages] = 0;
365 list_add(&m->list, &huge_boot_pages);
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700366 m->hstate = hstate;
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700367 return 1;
368}
Becky Bruce41151e72011-06-28 09:54:48 +0000369#endif
Jon Tollefsonec4b2c02008-07-23 21:27:53 -0700370
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800371int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
372{
373 return 0;
374}
375
Becky Bruce881fde12011-10-10 10:50:40 +0000376#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000377#define HUGEPD_FREELIST_SIZE \
378 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
379
380struct hugepd_freelist {
381 struct rcu_head rcu;
382 unsigned int index;
383 void *ptes[0];
384};
385
386static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
387
388static void hugepd_free_rcu_callback(struct rcu_head *head)
389{
390 struct hugepd_freelist *batch =
391 container_of(head, struct hugepd_freelist, rcu);
392 unsigned int i;
393
394 for (i = 0; i < batch->index; i++)
395 kmem_cache_free(hugepte_cache, batch->ptes[i]);
396
397 free_page((unsigned long)batch);
398}
399
400static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
401{
402 struct hugepd_freelist **batchp;
403
404 batchp = &__get_cpu_var(hugepd_freelist_cur);
405
406 if (atomic_read(&tlb->mm->mm_users) < 2 ||
407 cpumask_equal(mm_cpumask(tlb->mm),
408 cpumask_of(smp_processor_id()))) {
409 kmem_cache_free(hugepte_cache, hugepte);
410 return;
411 }
412
413 if (*batchp == NULL) {
414 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
415 (*batchp)->index = 0;
416 }
417
418 (*batchp)->ptes[(*batchp)->index++] = hugepte;
419 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
420 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
421 *batchp = NULL;
422 }
423}
424#endif
425
David Gibsona4fe3ce2009-10-26 19:24:31 +0000426static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
427 unsigned long start, unsigned long end,
428 unsigned long floor, unsigned long ceiling)
David Gibsonf10a04c2006-04-28 15:02:51 +1000429{
430 pte_t *hugepte = hugepd_page(*hpdp);
Becky Bruce41151e72011-06-28 09:54:48 +0000431 int i;
432
David Gibsona4fe3ce2009-10-26 19:24:31 +0000433 unsigned long pdmask = ~((1UL << pdshift) - 1);
Becky Bruce41151e72011-06-28 09:54:48 +0000434 unsigned int num_hugepd = 1;
435
Becky Bruce881fde12011-10-10 10:50:40 +0000436#ifdef CONFIG_PPC_FSL_BOOK3E
437 /* Note: On fsl the hpdp may be the first of several */
Becky Bruce41151e72011-06-28 09:54:48 +0000438 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
Becky Bruce881fde12011-10-10 10:50:40 +0000439#else
440 unsigned int shift = hugepd_shift(*hpdp);
Becky Bruce41151e72011-06-28 09:54:48 +0000441#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000442
443 start &= pdmask;
444 if (start < floor)
445 return;
446 if (ceiling) {
447 ceiling &= pdmask;
448 if (! ceiling)
449 return;
450 }
451 if (end - 1 > ceiling - 1)
452 return;
David Gibsonf10a04c2006-04-28 15:02:51 +1000453
Becky Bruce41151e72011-06-28 09:54:48 +0000454 for (i = 0; i < num_hugepd; i++, hpdp++)
455 hpdp->pd = 0;
456
David Gibsonf10a04c2006-04-28 15:02:51 +1000457 tlb->need_flush = 1;
Becky Bruce881fde12011-10-10 10:50:40 +0000458
459#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000460 hugepd_free(tlb, hugepte);
Becky Bruce881fde12011-10-10 10:50:40 +0000461#else
462 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
Becky Bruce41151e72011-06-28 09:54:48 +0000463#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000464}
465
David Gibsonf10a04c2006-04-28 15:02:51 +1000466static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
467 unsigned long addr, unsigned long end,
David Gibsona4fe3ce2009-10-26 19:24:31 +0000468 unsigned long floor, unsigned long ceiling)
David Gibsonf10a04c2006-04-28 15:02:51 +1000469{
470 pmd_t *pmd;
471 unsigned long next;
472 unsigned long start;
473
474 start = addr;
David Gibsonf10a04c2006-04-28 15:02:51 +1000475 do {
Becky Brucea1cd5412011-10-10 10:50:39 +0000476 pmd = pmd_offset(pud, addr);
David Gibsonf10a04c2006-04-28 15:02:51 +1000477 next = pmd_addr_end(addr, end);
478 if (pmd_none(*pmd))
479 continue;
Becky Brucea1cd5412011-10-10 10:50:39 +0000480#ifdef CONFIG_PPC_FSL_BOOK3E
481 /*
482 * Increment next by the size of the huge mapping since
483 * there may be more than one entry at this level for a
484 * single hugepage, but all of them point to
485 * the same kmem cache that holds the hugepte.
486 */
487 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
488#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000489 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
490 addr, next, floor, ceiling);
Becky Brucea1cd5412011-10-10 10:50:39 +0000491 } while (addr = next, addr != end);
David Gibsonf10a04c2006-04-28 15:02:51 +1000492
493 start &= PUD_MASK;
494 if (start < floor)
495 return;
496 if (ceiling) {
497 ceiling &= PUD_MASK;
498 if (!ceiling)
499 return;
500 }
501 if (end - 1 > ceiling - 1)
502 return;
503
504 pmd = pmd_offset(pud, start);
505 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000506 pmd_free_tlb(tlb, pmd, start);
David Gibsonf10a04c2006-04-28 15:02:51 +1000507}
David Gibsonf10a04c2006-04-28 15:02:51 +1000508
509static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
510 unsigned long addr, unsigned long end,
511 unsigned long floor, unsigned long ceiling)
512{
513 pud_t *pud;
514 unsigned long next;
515 unsigned long start;
516
517 start = addr;
David Gibsonf10a04c2006-04-28 15:02:51 +1000518 do {
Becky Brucea1cd5412011-10-10 10:50:39 +0000519 pud = pud_offset(pgd, addr);
David Gibsonf10a04c2006-04-28 15:02:51 +1000520 next = pud_addr_end(addr, end);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000521 if (!is_hugepd(pud)) {
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100522 if (pud_none_or_clear_bad(pud))
523 continue;
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700524 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
David Gibsona4fe3ce2009-10-26 19:24:31 +0000525 ceiling);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100526 } else {
Becky Brucea1cd5412011-10-10 10:50:39 +0000527#ifdef CONFIG_PPC_FSL_BOOK3E
528 /*
529 * Increment next by the size of the huge mapping since
530 * there may be more than one entry at this level for a
531 * single hugepage, but all of them point to
532 * the same kmem cache that holds the hugepte.
533 */
534 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
535#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000536 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
537 addr, next, floor, ceiling);
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100538 }
Becky Brucea1cd5412011-10-10 10:50:39 +0000539 } while (addr = next, addr != end);
David Gibsonf10a04c2006-04-28 15:02:51 +1000540
541 start &= PGDIR_MASK;
542 if (start < floor)
543 return;
544 if (ceiling) {
545 ceiling &= PGDIR_MASK;
546 if (!ceiling)
547 return;
548 }
549 if (end - 1 > ceiling - 1)
550 return;
551
552 pud = pud_offset(pgd, start);
553 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000554 pud_free_tlb(tlb, pud, start);
David Gibsonf10a04c2006-04-28 15:02:51 +1000555}
556
557/*
558 * This function frees user-level page tables of a process.
559 *
560 * Must be called with pagetable lock held.
561 */
Jan Beulich42b77722008-07-23 21:27:10 -0700562void hugetlb_free_pgd_range(struct mmu_gather *tlb,
David Gibsonf10a04c2006-04-28 15:02:51 +1000563 unsigned long addr, unsigned long end,
564 unsigned long floor, unsigned long ceiling)
565{
566 pgd_t *pgd;
567 unsigned long next;
David Gibsonf10a04c2006-04-28 15:02:51 +1000568
569 /*
David Gibsona4fe3ce2009-10-26 19:24:31 +0000570 * Because there are a number of different possible pagetable
571 * layouts for hugepage ranges, we limit knowledge of how
572 * things should be laid out to the allocation path
573 * (huge_pte_alloc(), above). Everything else works out the
574 * structure as it goes from information in the hugepd
575 * pointers. That means that we can't here use the
576 * optimization used in the normal page free_pgd_range(), of
577 * checking whether we're actually covering a large enough
578 * range to have to do anything at the top level of the walk
579 * instead of at the bottom.
David Gibsonf10a04c2006-04-28 15:02:51 +1000580 *
David Gibsona4fe3ce2009-10-26 19:24:31 +0000581 * To make sense of this, you should probably go read the big
582 * block comment at the top of the normal free_pgd_range(),
583 * too.
David Gibsonf10a04c2006-04-28 15:02:51 +1000584 */
585
David Gibsonf10a04c2006-04-28 15:02:51 +1000586 do {
David Gibsonf10a04c2006-04-28 15:02:51 +1000587 next = pgd_addr_end(addr, end);
Becky Bruce41151e72011-06-28 09:54:48 +0000588 pgd = pgd_offset(tlb->mm, addr);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000589 if (!is_hugepd(pgd)) {
David Gibson0b264252008-09-05 11:49:54 +1000590 if (pgd_none_or_clear_bad(pgd))
591 continue;
592 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
593 } else {
Becky Bruce881fde12011-10-10 10:50:40 +0000594#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000595 /*
596 * Increment next by the size of the huge mapping since
Becky Bruce881fde12011-10-10 10:50:40 +0000597 * there may be more than one entry at the pgd level
598 * for a single hugepage, but all of them point to the
599 * same kmem cache that holds the hugepte.
Becky Bruce41151e72011-06-28 09:54:48 +0000600 */
601 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
602#endif
David Gibsona4fe3ce2009-10-26 19:24:31 +0000603 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
604 addr, next, floor, ceiling);
David Gibson0b264252008-09-05 11:49:54 +1000605 }
Becky Bruce41151e72011-06-28 09:54:48 +0000606 } while (addr = next, addr != end);
David Gibsone28f7fa2005-08-05 19:39:06 +1000607}
608
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609struct page *
610follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
611{
612 pte_t *ptep;
613 struct page *page;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000614 unsigned shift;
615 unsigned long mask;
616
617 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700619 /* Verify it is a huge page else bail. */
David Gibsona4fe3ce2009-10-26 19:24:31 +0000620 if (!ptep || !shift)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 return ERR_PTR(-EINVAL);
622
David Gibsona4fe3ce2009-10-26 19:24:31 +0000623 mask = (1UL << shift) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 page = pte_page(*ptep);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000625 if (page)
626 page += (address & mask) / PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627
628 return page;
629}
630
631int pmd_huge(pmd_t pmd)
632{
633 return 0;
634}
635
Andi Kleenceb86872008-07-23 21:27:50 -0700636int pud_huge(pud_t pud)
637{
638 return 0;
639}
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641struct page *
642follow_huge_pmd(struct mm_struct *mm, unsigned long address,
643 pmd_t *pmd, int write)
644{
645 BUG();
646 return NULL;
647}
648
David Gibsona4fe3ce2009-10-26 19:24:31 +0000649static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
650 unsigned long end, int write, struct page **pages, int *nr)
651{
652 unsigned long mask;
653 unsigned long pte_end;
Andrea Arcangeli35267412011-11-02 13:37:15 -0700654 struct page *head, *page, *tail;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000655 pte_t pte;
656 int refs;
657
658 pte_end = (addr + sz) & ~(sz-1);
659 if (pte_end < end)
660 end = pte_end;
661
662 pte = *ptep;
663 mask = _PAGE_PRESENT | _PAGE_USER;
664 if (write)
665 mask |= _PAGE_RW;
666
667 if ((pte_val(pte) & mask) != mask)
668 return 0;
669
670 /* hugepages are never "special" */
671 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
672
673 refs = 0;
674 head = pte_page(pte);
675
676 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
Andrea Arcangeli35267412011-11-02 13:37:15 -0700677 tail = page;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000678 do {
679 VM_BUG_ON(compound_head(page) != head);
680 pages[*nr] = page;
681 (*nr)++;
682 page++;
683 refs++;
684 } while (addr += PAGE_SIZE, addr != end);
685
686 if (!page_cache_add_speculative(head, refs)) {
687 *nr -= refs;
688 return 0;
689 }
690
691 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
692 /* Could be optimized better */
Andrea Arcangeli85964682011-11-02 13:37:11 -0700693 *nr -= refs;
694 while (refs--)
Andrea Arcangeli405e44f2011-11-02 13:37:08 -0700695 put_page(head);
Andrea Arcangelicf592bf2011-11-02 13:37:19 -0700696 return 0;
697 }
698
699 /*
700 * Any tail page need their mapcount reference taken before we
701 * return.
702 */
703 while (refs--) {
704 if (PageTail(tail))
705 get_huge_page_tail(tail);
706 tail++;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000707 }
708
709 return 1;
710}
711
David Gibson39adfa52009-11-23 20:03:40 +0000712static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
713 unsigned long sz)
714{
715 unsigned long __boundary = (addr + sz) & ~(sz-1);
716 return (__boundary - 1 < end - 1) ? __boundary : end;
717}
718
David Gibsona4fe3ce2009-10-26 19:24:31 +0000719int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
720 unsigned long addr, unsigned long end,
721 int write, struct page **pages, int *nr)
722{
723 pte_t *ptep;
724 unsigned long sz = 1UL << hugepd_shift(*hugepd);
David Gibson39adfa52009-11-23 20:03:40 +0000725 unsigned long next;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000726
727 ptep = hugepte_offset(hugepd, addr, pdshift);
728 do {
David Gibson39adfa52009-11-23 20:03:40 +0000729 next = hugepte_addr_end(addr, end, sz);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000730 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
731 return 0;
David Gibson39adfa52009-11-23 20:03:40 +0000732 } while (ptep++, addr = next, addr != end);
David Gibsona4fe3ce2009-10-26 19:24:31 +0000733
734 return 1;
735}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
Becky Bruce76512952011-10-10 10:50:36 +0000737#ifdef CONFIG_PPC_MM_SLICES
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
739 unsigned long len, unsigned long pgoff,
740 unsigned long flags)
741{
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700742 struct hstate *hstate = hstate_file(file);
743 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
Brian King48f797d2008-12-04 04:07:54 +0000744
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700745 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746}
Becky Bruce76512952011-10-10 10:50:36 +0000747#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Mel Gorman33402892009-01-06 14:38:54 -0800749unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
750{
Paul Mackerras25c29f92011-09-20 19:58:10 +0000751#ifdef CONFIG_PPC_MM_SLICES
Mel Gorman33402892009-01-06 14:38:54 -0800752 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
753
754 return 1UL << mmu_psize_to_shift(psize);
Becky Bruce41151e72011-06-28 09:54:48 +0000755#else
756 if (!is_vm_hugetlb_page(vma))
757 return PAGE_SIZE;
758
759 return huge_page_size(hstate_vma(vma));
760#endif
761}
762
763static inline bool is_power_of_4(unsigned long x)
764{
765 if (is_power_of_2(x))
766 return (__ilog2(x) % 2) ? false : true;
767 return false;
Mel Gorman33402892009-01-06 14:38:54 -0800768}
769
David Gibsond1837cb2009-10-26 19:24:31 +0000770static int __init add_huge_page_size(unsigned long long size)
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100771{
David Gibsond1837cb2009-10-26 19:24:31 +0000772 int shift = __ffs(size);
773 int mmu_psize;
David Gibsona4fe3ce2009-10-26 19:24:31 +0000774
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100775 /* Check that it is a page size supported by the hardware and
David Gibsond1837cb2009-10-26 19:24:31 +0000776 * that it fits within pagetable and slice limits. */
Becky Bruce41151e72011-06-28 09:54:48 +0000777#ifdef CONFIG_PPC_FSL_BOOK3E
778 if ((size < PAGE_SIZE) || !is_power_of_4(size))
779 return -EINVAL;
780#else
David Gibsond1837cb2009-10-26 19:24:31 +0000781 if (!is_power_of_2(size)
782 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
783 return -EINVAL;
Becky Bruce41151e72011-06-28 09:54:48 +0000784#endif
Jon Tollefson91224342008-07-23 21:27:55 -0700785
David Gibsond1837cb2009-10-26 19:24:31 +0000786 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
787 return -EINVAL;
788
789#ifdef CONFIG_SPU_FS_64K_LS
790 /* Disable support for 64K huge pages when 64K SPU local store
791 * support is enabled as the current implementation conflicts.
792 */
793 if (shift == PAGE_SHIFT_64K)
794 return -EINVAL;
795#endif /* CONFIG_SPU_FS_64K_LS */
796
797 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
798
799 /* Return if huge page size has already been setup */
800 if (size_to_hstate(size))
801 return 0;
802
803 hugetlb_add_hstate(shift - PAGE_SHIFT);
804
805 return 0;
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100806}
807
808static int __init hugepage_setup_sz(char *str)
809{
810 unsigned long long size;
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100811
812 size = memparse(str, &str);
813
David Gibsond1837cb2009-10-26 19:24:31 +0000814 if (add_huge_page_size(size) != 0)
Jon Tollefson4ec161c2008-01-04 09:59:50 +1100815 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
816
817 return 1;
818}
819__setup("hugepagesz=", hugepage_setup_sz);
820
Becky Bruce881fde12011-10-10 10:50:40 +0000821#ifdef CONFIG_PPC_FSL_BOOK3E
Becky Bruce41151e72011-06-28 09:54:48 +0000822struct kmem_cache *hugepte_cache;
823static int __init hugetlbpage_init(void)
824{
825 int psize;
826
827 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
828 unsigned shift;
829
830 if (!mmu_psize_defs[psize].shift)
831 continue;
832
833 shift = mmu_psize_to_shift(psize);
834
835 /* Don't treat normal page sizes as huge... */
836 if (shift != PAGE_SHIFT)
837 if (add_huge_page_size(1ULL << shift) < 0)
838 continue;
839 }
840
841 /*
842 * Create a kmem cache for hugeptes. The bottom bits in the pte have
843 * size information encoded in them, so align them to allow this
844 */
845 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
846 HUGEPD_SHIFT_MASK + 1, 0, NULL);
847 if (hugepte_cache == NULL)
848 panic("%s: Unable to create kmem cache for hugeptes\n",
849 __func__);
850
851 /* Default hpage size = 4M */
852 if (mmu_psize_defs[MMU_PAGE_4M].shift)
853 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
854 else
855 panic("%s: Unable to set default huge page size\n", __func__);
856
857
858 return 0;
859}
860#else
David Gibsonf10a04c2006-04-28 15:02:51 +1000861static int __init hugetlbpage_init(void)
862{
David Gibsona4fe3ce2009-10-26 19:24:31 +0000863 int psize;
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700864
Matt Evans44ae3ab2011-04-06 19:48:50 +0000865 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
David Gibsonf10a04c2006-04-28 15:02:51 +1000866 return -ENODEV;
Benjamin Herrenschmidt00df4382008-07-28 16:13:18 +1000867
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700868 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
David Gibsond1837cb2009-10-26 19:24:31 +0000869 unsigned shift;
870 unsigned pdshift;
871
872 if (!mmu_psize_defs[psize].shift)
873 continue;
874
875 shift = mmu_psize_to_shift(psize);
876
877 if (add_huge_page_size(1ULL << shift) < 0)
878 continue;
879
880 if (shift < PMD_SHIFT)
881 pdshift = PMD_SHIFT;
882 else if (shift < PUD_SHIFT)
883 pdshift = PUD_SHIFT;
884 else
885 pdshift = PGDIR_SHIFT;
886
887 pgtable_cache_add(pdshift - shift, NULL);
888 if (!PGT_CACHE(pdshift - shift))
889 panic("hugetlbpage_init(): could not create "
890 "pgtable cache for %d bit pagesize\n", shift);
Jon Tollefson0d9ea752008-07-23 21:27:56 -0700891 }
David Gibsonf10a04c2006-04-28 15:02:51 +1000892
David Gibsond1837cb2009-10-26 19:24:31 +0000893 /* Set default large page size. Currently, we pick 16M or 1M
894 * depending on what is available
895 */
896 if (mmu_psize_defs[MMU_PAGE_16M].shift)
897 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
898 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
899 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
900
David Gibsonf10a04c2006-04-28 15:02:51 +1000901 return 0;
902}
Becky Bruce41151e72011-06-28 09:54:48 +0000903#endif
David Gibsonf10a04c2006-04-28 15:02:51 +1000904module_init(hugetlbpage_init);
David Gibson0895ecd2009-10-26 19:24:31 +0000905
906void flush_dcache_icache_hugepage(struct page *page)
907{
908 int i;
Becky Bruce41151e72011-06-28 09:54:48 +0000909 void *start;
David Gibson0895ecd2009-10-26 19:24:31 +0000910
911 BUG_ON(!PageCompound(page));
912
Becky Bruce41151e72011-06-28 09:54:48 +0000913 for (i = 0; i < (1UL << compound_order(page)); i++) {
914 if (!PageHighMem(page)) {
915 __flush_dcache_icache(page_address(page+i));
916 } else {
Cong Wang2480b202011-11-25 23:14:16 +0800917 start = kmap_atomic(page+i);
Becky Bruce41151e72011-06-28 09:54:48 +0000918 __flush_dcache_icache(start);
Cong Wang2480b202011-11-25 23:14:16 +0800919 kunmap_atomic(start);
Becky Bruce41151e72011-06-28 09:54:48 +0000920 }
921 }
David Gibson0895ecd2009-10-26 19:24:31 +0000922}